Shape and Size from the Mist: A Deformable Model for Particle Characterization

Process optimization often depends on the correct estimation of particle size, their shape and their concentration. In case of the backlight microscopic system, which we investigate here, particle images suffer from out-of-focus blur. This gives a bias towards overestimating the particle size when particles are behind or in front of the focus plane. In most applications only in-focus particles get analyzed, but this weakens the statistical basis and requires either particle sampling over longer time or results in uncertain predictions. We propose a new method for estimating the size and the shape of the particles, which includes out-of-focus particles. We employ particle simulations for training an inference model predicting the true size of particles from image observations. This also provides depth information, which can be used in concentration predictions. Our model shows promising results on real data with ground truth depth, shape and size information. The outcome of our approach is a reliable particle analysis obtained from shorter sampling time.